Large scale graph processing systems: survey and an experimental - PowerPoint PPT Presentation

Large scale graph processing systems: survey and an experimental evaluation Cluster Computing 2015 Omar Batarfi , Radwa El Shawi, Ayman G. Fayoumi , Reza Nouri, Seyed-Mehdi-Reza Beheshti, Ahmed Barnawi, Sherif Sakr

Graph scale  Scale: millions to billions of nodes and edges  Facebook social network graph:  1 billion+ users (nodes)  140 billion+ friendship relationships (edges)  The size of a single node/edge can be very different due to the various attributes size or nodes/edges  An Estimation of large-scale graph size: 1 GB to 500 TB  ~0.5 GB/million nodes, ~0.20 GB/million edges  1 million nodes/edges: ~0.5 GB nodes & ~0.20 GB edges  500 billion nodes/edges: ~250 TB nodes & ~100 TB edges GB/million nodes GB/million edges 0.42 0.20 0.84 0.13 11.39 2.60 1.83 0.07 0.58 0.23 0.48 0.19 0.44 0.18

platforms  General-purpose platforms (such as MapReduce) are bad  No direct support for iterative graph algorithms  To detect a fix point (termination condition), an extra task might be required on each iteration  Specialized platforms  Pregel family  Graphlab family  Others

Algorithms  Characteristics of graph algorithms  Iterative  Need to traverse the graph  Typical graph algorithms  PageRank : rank nodes based on their incoming/outgoing edges  Shortest Path : find the path between 2 nodes where the sum of weights is minimal  Pattern Matching : find certain structures (e.g. path, star)  Triangle Count : counts the number of triangles  Connected Component : find the subgraphs in which any two vertices are connected

Pregel family  Pregel:  Google’s pioneer work in this area  Published in 2010  Distributed & computations are totally in memory  Iteration -> superstep  Address scalability issue  Bulk Synchronous Parallel (BSP ): synchronization barrier on each superstep  Message passing interface (MPI)  Vertex-centric approach  Locality : Each vertex & its neighbors are in the same node  A vertex can: execute a function/send messages to others/change states (active/inactive)  Termination: no active vertices & no messages being transmitted  Pregel family  Apache Giraph: Java implementation of pregel  GPS: another Java implementation  Pregelix: set- oriented, iterative dataflow

GraphLab family  GraphLab  Shared memory  GAS (Gather, Apply, Scatter) processing model  Gather: a vertex collects info of its neighbors  Apply: performs computation  Scatter: update adjacent vertices and edges  Comparison  GAS : pull-based; a vertex request info of all neighbors  MPI: push-base; a vertex receives info from neighbors  Two modes:  Synchronous model (BSP): communication barriers  Asynchronous model: using distributed locking; no communication barriers orsuperstep  GraphLab family  PowerGraph: avoid the imbalanced workload caused by high degree vertices in power-law graphs  Trinity: memory-based; distributed  Signal/Collect: vertex-centric; two operations for a vertex (signal/collect)  Graphchi

GraphLab family cont’d  Graphchi :  Out-of-core: using secondary storage in a single machine  Parallel Sliding Window (PSW):  Goal: decreases non-sequential accesses on disk  It partitions the graph into shards  In each shard, edges are sorted by the source IDs  Selective scheduling:  Converge faster on some parts of the graph  “some parts” - > the change on values is significant  Pros  It avoids the challenge of finding efficient graph cuts  Now with zone-based devices, partitioning is needed again  It avoids cluster management, fault tolerance etc.  Out-of-Core + SMR

Other systems  TurboGraph  Out-of-core  Processing billion-scale graphs using modern hardware -> parallelism  Multicore CPU: multiple job at the same time  FlashSSD : multiple I/O requests in parallel using multiple flash memory packages  A parallel model called pin-and-slide : column view of the matrix-vector multiplication  Two types of thread pools  Execution thread pool  Asynchronous I/O callback thread pool  Steps  Restrict computation to a set of vertice -> identify the corresponding pages  -> pin those pages in the buffer pool  -> processing completes for a page -> swtich unpinned -> can be evicted now  -> Parallel asynchronous I/O request to the FlashSSD for pages which are not in the buffer pool  The system can slide the processing window one page at a time  Multiple channel SSD  Extreme-large-scale graph that does fit into memory  CMR -> SMR/Zone named SSD

Other systems  GRACE  Out-of-core  Batch-style graph programming frameworks  Providing a high level representation for graph data  Separating application logic from execution policies.  Combine synchronous programming with asynchronous execution

Experiments  Perforamance metrics  Reading Time , Processing Time, Writing Time, Total Execution Time, CPU Utilization, RAM Usage, Network Traffic  Deployed on Amazon AWS cloud services The execution times metrics for the PrageRank algorithm for all systems using the different datasets